Evidence
No promises. Figures.
Everything we claim on this site is backed by independent research and official statistics. That includes the figures that are less flattering to AI. So you can judge for yourself.
- Belgium
- EU (27)
32 sources from Eurostat, the European Union, the Belgian government, universities and research institutes.
Last checked on 25 September 2026. Every figure links to the original source.
Honestly labelled. We show the limitations of every study and say when something is our own calculation or an assumption.
Where does developers' time go?
Most of a developer's work is not writing new code. A large share goes to maintenance, searching and work that keeps getting postponed.
- 23.8 h other work
- 13.5 h technical debt
- 3.8 h bad code
Together 17.3 of 41.1 hours, or 42% of the week (our own calculation). Survey from 2018 in five countries, not including Belgium.
Source: Stripe and Harris Poll (2018) [5]
Show the figures as a table
| Hours per week | |
|---|---|
| other work | 23.8 h |
| technical debt | 13.5 h |
| bad code | 3.8 h |
Self-reported by 3,500 developers and managers in six countries.
Source: Atlassian (2025) [6]
Show the figures as a table
| Share | |
|---|---|
| lose 6 hours or more per week | 90% |
| lose 10 hours or more per week | 50% |
By default, our calculator assumes 8 hours of postponed work per developer per week. That is deliberately less than what developers report themselves in these studies. We would rather be cautious than make promises.
Technical debt costs money every week
Old or hasty solutions from the past slow down everything that follows. Large consultancies and research firms have mapped that cost.
McKinsey surveyed CIOs of large financial and technology companies. CISQ made an estimate for the United States. No comparable figures exist for Belgian SMEs. The direction is the same, though.
Reducing technical debt in small, tested steps is one of the use cases where agents are strongest. The work is dull, well defined and easy to check.
Reducing technical debtWhat does AI deliver in programming?
Randomised studies exist, the strictest kind of research. Their results vary widely. We would rather tell you that ourselves.
Effect in per cent
The studies do not measure the same thing (time per task or number of completed tasks) and are not directly comparable. The margin of error of the field experiment is calculated from the reported standard error (our own calculation). METR now calls its 2025 result outdated; according to METR itself, a follow-up measurement in 2026 did not yet produce a reliable new figure [15].
Source: Peng, Kalliamvakou, Cihon and Demirer (2023) [11] · Cui, Demirer, Jaffe, Musolff, Peng and Salz (2026) [13] · Paradis, Chandra et al. (Google) (2024) [12] · METR (2025) [14]
Show the figures as a table
| Effect | Margin of error (95%) | |
|---|---|---|
| GitHub Copilot, lab experiment (2023) | +55.8% | 21.0% to 89.0% |
| Field experiment at 3 companies (2026) | +26.1% | 5.9% to 46.3% |
| Google, randomised study (2024) | +21.0% | not reported |
| METR, experienced developers (2025) | -19.0% | -39.0% to -2.0% |
What works
Well-defined tasks with a clear end point deliver the biggest gains. In the largest field experiment, less experienced developers gained the most, according to the researchers. [11, 13]
What does not work by itself
Experienced developers on large, complex projects became slower with AI assistance in 2025. Without a good way of working, AI sometimes costs more time than it saves. [14]
The effect depends on the task, the experience and the way of working. That is why we do not promise a fixed percentage. We choose the tasks where agents are strong and measure the result in your business.
Perception is not measurement
In the METR study, developers thought AI made them faster. The clock said otherwise.
Effect on speed in per cent
Source: METR (2025) [14]
Show the figures as a table
| Effect | Margin of error (95%) | |
|---|---|---|
| Expectation beforehand | +24.0% | not reported |
| Perception afterwards | +20.0% | not reported |
| Measured by the clock | -19.0% | -39.0% to -2.0% |
Both figures come from the same independent research institute. They show how hard it is to estimate time saved by gut feeling.
That is why we measure time instead of asking how it feels. Every report shows the hours won back and the review time.
How we measureTrust and security
Developers use AI on a massive scale, but they do not trust the output blindly. Rightly so.
- 3.1% highly trust
- 29.6% somewhat trust
- 21.6% other answers
- 26.1% somewhat distrust
- 19.6% highly distrust
In total, 32.7% trust the output and 45.7% distrust it. "Other answers" is the difference to 100% (our own calculation).
Source: Stack Overflow (2025) [20]
Show the figures as a table
| Share | |
|---|---|
| highly trust | 3.1% |
| somewhat trust | 29.6% |
| other answers | 21.6% |
| somewhat distrust | 26.1% |
| highly distrust | 19.6% |
Source: Stack Overflow (2025) [20]
Show the figures as a table
| Share | |
|---|---|
| solutions that are almost right, but not quite | 66.0% |
| debugging AI code takes more time | 45.2% |
Veracode sells security scanners and gave the models no instructions about security. DORA works with surveys and statistical models, not experiments. DORA sums it up itself: AI amplifies what is already there in a team, for better and for worse.
That is why, with us, a human approves every change, agents work in an isolated environment and the tests run on every proposal.
How we work securelyWhy now?
AI agents can complete ever longer tasks on their own. The pace of that progress has been measured.
METR measures the task length (in human working time) that an agent completes with a 50% success rate, mainly for software tasks. A longer task length does not yet mean productivity gains in practice. That requires a good way of working.
The technology is usable for well-defined work. The challenge now lies in the way of working: which tasks, which rules, who reviews. That is our job.
Belgium in figures
Belgian businesses are ahead of the European average. But know-how remains the biggest obstacle.
From 10.3% in 2021 to 34.5% in 2025. AI use in Belgian businesses has more than tripled in four years and is well above the European average of 20%. [1]
The exact figures per year and per company size are in the charts below.
- Belgium
- European Union (27)
There was no measurement in 2022. In 2021, Eurostat used a slightly different list of AI technologies.
Source: Eurostat (2025) [1]
Show the figures as a table
| Belgium | European Union (27) | |
|---|---|---|
| 2021 | 10.3% | 7.7% |
| 2023 | 13.8% | 8.1% |
| 2024 | 24.7% | 13.5% |
| 2025 | 34.5% | 20.0% |
- Belgium
- European Union (27)
Source: Eurostat (2025) [1]
Show the figures as a table
| Belgium | European Union (27) | |
|---|---|---|
| 10 to 49 employees | 28.8% | 17.0% |
| 50 to 249 employees | 54.5% | 30.4% |
| 250 or more | 76.4% | 55.0% |
Lack of expertise is also the main reason across the EU (70.3%). Among the smallest Belgian businesses (10 to 49 employees) it is 74.9%.
Source: Eurostat (2025) [1]
Show the figures as a table
| Share | |
|---|---|
| Lack of relevant expertise | 74.5% |
| Unclear legal consequences | 57.9% |
| Data protection concerns | 50.9% |
| Availability or quality of data | 46.2% |
| Costs too high | 35.5% |
Our own calculation: 9.07% of all businesses had hard-to-fill ICT vacancies, out of 14.50% that were recruiting ICT specialists. The EU figures are 5.49% and 9.55%.
Three in four Belgian businesses that considered AI give up because of a lack of expertise. And finding extra ICT staff is hard. We bring that expertise, without you having to hire anyone.
Rules and support
The EU AI Act is in force and was amended in July 2026. This is where things stand.
The EU AI Act enters into force. [29]
Prohibited AI practices and the AI literacy obligation (Article 4) apply. [29]
Rules for general-purpose AI models apply. [29]
General application, including the transparency obligations of Article 50. [29, 30]
Rules for high-risk systems listed in Annex III, postponed by the AI omnibus. [30]
Rules for high-risk systems in regulated products (Annex I), postponed by the AI omnibus. [30]
AI literacy (Article 4)
Businesses that use AI systems must take measures to support the AI literacy of their staff. Since the AI omnibus, the text says "support" instead of "ensure". [30]
According to the European Commission, no certificate is required. It recommends keeping an internal record of the training given, and points out that simply having staff read manuals has little effect. [31]
SME wallet (KMO-portefeuille, Flanders)
For training, small enterprises receive 30% support and medium-sized ones 20%, up to a maximum of €7,500 per year. Online training qualifies if there is direct interaction with the trainer. The training must be delivered by a registered service provider. [32]
Since 1 February 2026, advice only qualifies if it concerns cybersecurity. [32]
Time-Regained is not yet a registered service provider. We will say so here as soon as it is.
This is general information, not legal advice. If in doubt, consult a lawyer.
Our calculation assumptions
The calculator and the worked example on this site use these values. For each value, you can see whether it comes from a source, is our own calculation or is an assumption.
| Value | What we use | Basis | Type |
|---|---|---|---|
| Working week | 38 hours | Statutory working hours in Belgium since 2003. [22] | Source |
| Working weeks per year | 46 | 52 weeks minus 4 weeks of holiday (20 days) and 2 weeks of public holidays (10 days). Sickness and extra leave not included. [24, 23] | Own calculation |
| Working day | 7.6 hours | 38 hours divided by 5 days. | Own calculation |
| Postponed work | 8 hours per developer per week | Deliberately lower than the 17.3 hours of maintenance from the Stripe study and the 10 hours or more that half of developers lose according to Atlassian. [5, 6] | Cautious assumption |
| Share an agent takes over | 50% | Depends heavily on the tasks. We measure this in the scan and the trial. | Assumption |
| Review time | 25% of the work taken over | Depends on the size of the proposals and the experience of your team. We measure this during implementation. | Assumption |
| Cost of one hour of developer time | €65 | Median gross salary of a software developer €4,652 per month, senior €4,000 to €5,500 per month. With the end-of-year bonus, holiday pay and 25% employer contributions, that comes to about €47 per hour (median) to €55 (senior). With fringe benefits and overheads, about €60 to €65. For comparison: the average labour cost in Belgium across all sectors was €48.2 per hour in 2024. [26, 25, 27, 28] | Own calculation |
You can adjust all assumptions yourself in the calculator.
All sources
Last checked on 25 September 2026. Spotted a mistake or a more recent source? Let us know.
-
1
Eurostat, Use of artificial intelligence in enterprises (datasets isoc_eb_ai and isoc_eb_ain2), data 2021 to 2025 · December 2025 Enterprises with 10 or more employees, excluding agriculture, mining and the financial sector. Data updated on 15 June 2026. No measurement in 2022; 2021 used a slightly different list of AI technologies.
-
2
Eurostat, 20% of EU enterprises use AI technologies (news release) · December 2025
-
3
Eurostat, Enterprises that recruited or tried to recruit ICT specialists (isoc_ske_itrcrn2), data 2024 · June 2026 Enterprises with 10 or more employees. This question is asked every two years; 2024 is the most recent year.
-
4
Eurostat, Number of ICT specialists in the EU continues to grow (news release, data 2025) · May 2026
-
5
Stripe and Harris Poll, The Developer Coefficient · September 2018 Survey of more than 1,000 developers and 1,000 managers in the US, the UK, France, Germany and Singapore. Self-estimated time.
-
6
Atlassian, State of Developer Experience Report 2025 · July 2025 3,500 developers and managers in six countries. Self-reported time.
-
7
Atlassian and DX, State of Developer Experience Report 2024 · July 2024 More than 2,100 developers and managers. Self-reported time.
-
8
McKinsey, Tech debt: Reclaiming tech equity · October 2020 Survey of 50 CIOs at financial and technology companies with more than 1 billion dollars in revenue.
-
9
McKinsey, Breaking technical debt's vicious cycle to modernize your business · April 2023
-
10
CISQ, The Cost of Poor Software Quality in the US: A 2022 Report · November 2022 Macroeconomic estimate for the United States.
-
11
Peng, Kalliamvakou, Cihon and Demirer, The Impact of AI on Developer Productivity: Evidence from GitHub Copilot · February 2023 Controlled experiment with developers recruited through Upwork. One well-defined task: writing an HTTP server in JavaScript. Authors affiliated with GitHub and Microsoft.
-
12
Paradis, Chandra et al. (Google), How much does AI impact development speed? An enterprise-based randomized controlled trial · October 2024 96 full-time Google engineers, one task with internal tools. The authors point to a wide margin of error.
-
13
Cui, Demirer, Jaffe, Musolff, Peng and Salz, The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers (Management Science) · February 2026 4,867 developers at Microsoft, Accenture and a Fortune 100 company. Measures the number of completed tasks, not their quality.
-
14
METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity · July 2025 16 experienced developers, 246 real tasks in large open-source projects. Margin of error: 2% to 39% more time.
-
15
METR, Uplift update (follow-up study) · February 2026 METR calls the 2025 result outdated and the new measurement unreliable because of task selection.
- 16
-
17
METR, Measuring AI Ability to Complete Long Tasks (2025) and Time Horizon 1.1 (2026) · January 2026 Task length (in human working time) that an agent completes with a 50% success rate, mainly software tasks.
-
18
Google Cloud DORA, Accelerate State of DevOps Report 2024 · October 2024 Survey; relationships are statistical models, not experiments.
-
19
Google Cloud DORA, State of AI-assisted Software Development 2025 · September 2025 Almost 5,000 respondents.
-
20
Stack Overflow, Developer Survey 2025: AI · July 2025 48,936 responses; the AI questions were answered by 68.7%.
-
21
Veracode, 2026 GenAI Code Security Report · July 2026 Research by a vendor of security scanners. Tasks without instructions about security.
-
22
FPS Employment, Labour and Social Dialogue, Arbeidsduur en vermindering van arbeidsduur (Working hours and reduction of working hours) · 2026
-
23
FPS Employment, Labour and Social Dialogue, Feestdagen (Public holidays) · 2026
-
24
National Annual Holiday Office (RJV), Berekening vakantiedagen (Calculation of holiday entitlement) · 2026
-
25
Hays Belgium, Gemiddeld IT-salaris in België 2026 (Average IT salary in Belgium 2026) · 2026 Gross monthly salaries.
-
26
Robert Half Belgium, Salaris software developer (Software developer salary) · 2026 National salary scale, gross per month.
-
27
VBO-FEB, Sociale bijdragen derde kwartaal 2026 (Social security contributions, third quarter 2026) · July 2026 Basic employer contribution in the private sector: 25%.
- 28
-
29
European Union, Regulation (EU) 2024/1689 (EU AI Act) · July 2024
-
30
European Union, Regulation (EU) 2026/1744 (AI omnibus, amending the EU AI Act) · July 2026 In force since 27 July 2026.
-
31
European Commission, AI literacy: questions and answers · July 2026
-
32
VLAIO, Kmo-portefeuille: hoeveel subsidie en voorwaarden (SME wallet: subsidy amounts and conditions) · February 2026 Only for Flemish SMEs.
Would you rather have your own figures?
Research shows averages. In a free scan, we look at your code, your tasks and your team, and you get an honest estimate.